Particle Level Noise Removal Using Machine Learning and Graph Neural Networks: A Comparative Study

Particle physics has always been an intriguing topic for scientists in the research field. Exotic particles pave the way for the exploration of the early universe. To separate exotic particles from the noise, scientists have done much research in the field of particle-level noise removal. This study addresses the problem of particle-level noise removal. Traditional machine learning models have been able to mitigate particle-level noise but they are inefficient in terms of accuracy. The introduction of Graph Neural Network and its various models addresses the efficiency of particle-level noise removal. This research compares the traditional machine learning model, specifically XGBoost and Support Vector Machine(SVM) with Graph Attention Network(GAT) to find the difference in efficiency between them. The data which was used to compare these models were generated using PYTHIA 8.3. The models were hyper-tuned and trained on 10K proton-proton collisions. Where GAT has achieved 80.1% accuracy, while XGBoost and SVM have achieved 72.9% and 67.3% accuracy. After training the model it was evaluated using Explainable AI (XAI). XAI GraphLIME was used for node-specific explanation and SHAP for feature importance. XAI-enhanced GAT achieved 94.8% accuracy, outperforming traditional methods.